low-cost wireless sensor networks for remote cardiac patients monitoring applications

WIRELESS COMMUNICATIONS AND MOBILE COMPUTINGWirel. Commun. Mob. Comput. 2008; 8:513–529Published online 29 January 2007 in Wiley InterScience(www.interscience.wiley.com) DOI: 10.1002/wcm.488

Low-cost wireless sensor networks for remote cardiacpatients monitoring applications

Fei Hu1, Meng Jiang1 and Yang Xiao2∗,†1Computer Engineering Department, Rochester Institute of Technology, Rochester, NY, U.S.A.2Department of Computer Science, University of Alabama, Tuscaloosa, AL, U.S.A.

Summary

One of today’s most pressing matters in medical care is response time to patients in need. Scope of this researchis to suggest a solution that would help reduce response time in emergency situations utilizing technologies ofwireless sensor networks. The enhanced power efficiency, minimized production cost, condensed physical layout,and reduced wired connections present a much more proficient and simplified approach to the continuous monitoringof patients’ physiological status. The proposed sensor network system is composed of wearable vital sign sensorsand a workstation monitor. The wearable platforms are to be distributed to patients of concern. The wearableplatforms can provide continuous electrocardiogram (ECG) monitoring by measuring electrical potentials betweenvarious points of the body using a galvanometer. They will then relay the ECG signals wirelessly to the workstationmonitor. In addition to displaying the data, the workstation will also perform signal wavelet transformation for ECGcharacteristic extractions. Copyright © 2007 John Wiley & Sons, Ltd.

KEY WORDS: tele-cardiology; wireless sensor networks; cardiac monitoring

1. Introduction

Recently, the rapid development of mobile technolo-gies, including increased communication bandwidthand miniaturization of mobile terminals, has acceler-ated developments in the field of mobile telemedicine.There are many researches on mobile telemedicine [1–25]. Development of efficient portable, body-worn de-vices and systems to measure physiological parametersof cardiac patients (e.g., ECG, heart rate, Pulse Oxime-try (SpO2), and blood pressure) has been witnessingactive research efforts lately. The implications and po-tentials of these wearable health monitoring technolo-gies are paramount, for their abilities to:(1) detect early

*Correspondence to: Yang Xiao, Department of Computer Science, University of Alabama, Tuscaloosa, AL, U.S.A.†E-mail: [email protected]

signs of health deterioration, (2) notify health careproviders in critical situations, (3) find correlations be-tween lifestyle and health issues, (4) bring sports con-ditioning into a new dimension, by providing detailedinformation about physiological signals under variousexercise conditions, and (5) transform health care byproviding doctors with multi-sourced, real-time phys-iological data.

ECG monitoring and interpretation have alwaysbeen tasks conventionally assigned to trained medicalcare personals. Although being more comprehensive inthe related knowledge, the constraints to manpower arealso very obvious. Fatigue factors and overwhelmingworkloads are both possible causes to delayed emer-

Copyright © 2007 John Wiley & Sons, Ltd.

514 F. HU, M. JIANG AND Y. XIAO

gency response that may have reduced the chances forpatients’ survival. By automating this process, the sys-tem frees the medical professionals of the tedious tasksto center their attentions on something much more de-manding.

There is still a significant gap between the existingsensor network solutions and the needs in medical care.This research is an endeavor to help closing the gap bysuggesting a solution for an ECG vital sign monitoringsystem designed to reduce the medical response timefor the patients in need. This monitoring system pro-vides many useful applications, giving support to thecurrent medical care structure, especially for patientsin intensive care units (ICUs) or under emergency re-lief situations. Under these scenarios, the physiologicalstatuses of multiple patients are continuously observedfor immediate medical decisions that may well increasetheir chances of survival.

Based on these motivations, there have been numeralattempts to develop medical systems similar to the pro-posed work in this research. Such efforts are primarilyled by the academia but extending deeply into the in-dustries. However, most research efforts have been fo-cusing on either the vital sign monitoring aspect or theECG feature extraction using standard databases, bothfalling short of expectation. Having analyzed the exist-ing solutions, this research will bridge the two majorresearch efforts and bring out a more realizable prod-uct to directly benefit the consumers in the medicalfield.

This research offers the following contributions tothe proposed system: foremost is the wearable ECGmonitoring platform, presented in Section 3, based ona 3-Lead System and a design under the CodeBlueproject [1]. The ECG data collected using these mobileplatforms are then transmitted wirelessly using TmoteSky via radio frequencies to a receiving mote connectedto the workstation monitor. The received patient dataon the workstation are processed using wavelet trans-forms [10,14] to provide feature extraction capabilities[7] in order to locate the characteristic points of theECG waves, shown in Section 5.

The rest of this paper is organized as follows. Sec-tion 2 provides some background information on thetopics of ECG signals, wavelet transformation the-ories, as well as others works supporting this re-search. Sections 3–5 offer more insights regarding theworks behind the wearable platform, wireless com-munication, and signal processing, respectively. Sec-tion 6 provides the final remarks of this work, futureoutlook, and recommendations for some subsequentwork.

2. Background Knowledge

Here we will provide a brief introduction to the subjectsof ECG interpretation, wireless sensor networks, andthe supporting work environment. It is by no means athorough tutorial for any of the above topics, but merelyan attempt to provide some basic knowledge necessaryto understand the underlying research work that wentinto preparing for this research.

2.1. ECG Interpretation

ECG [11] is abbreviated from the word electrocardio-gram, or alternatively called EKG, which is the ab-breviation of the German word elektrokardiogramm.Produced by an electrocardiograph, the signal is con-structed by measuring electrical potentials betweenvarious points of the body using a galvanometer. ECGsignals have a wide array of applications throughoutthe medical field in determining whether the heart isfunctioning properly or suffering from any abnormal-ities. It helps to screen and diagnosis cardiovasculardiseases such as ischemia, cardiac arrhythmia, and mi-tral stenosis, etc.

Figure 1 shows an example of a normal ECG trace,which consists of a P wave, a QRS complex, and a Twave. A small U wave may also be sometimes visi-ble, but is neglected in this work for its inconsistency.The P wave is the electrical signature of the currentthat causes atrial contraction; the QRS complex corre-sponds to the current that causes contraction of the leftand right ventricles; the T wave represents the repolar-ization of the ventricles; and the U wave, although notalways visible, is considered to be a representation ofthe papillary muscles or Purkinje fibers [12]. The pres-ence or lack of presence of these waves as well as theQT interval and PR interval are meaningful parameters

Fig. 1. Example of a normal ECG trace.

Copyright © 2007 John Wiley & Sons, Ltd. Wirel. Commun. Mob. Comput. 2008; 8:513–529

DOI: 10.1002/wcm

LOW-COST WSN FOR REMOTE MONITORING APPLICATIONS 515

Fig. 2. Electrocardiogram of a healthy heart.

in the screening and diagnosis of cardiovascular dis-eases. Figure 2 is an example of a healthy ECG whichshows clearly all of the components mentioned above.

There are several flavors of the system used to mon-itor patient ECG information differing primarily onlead placements, ranging from 3-leads to 12-leads. The3-lead system is non-diagnostic and is meant forrhythm interpretation, while the 12-lead system, on theother hand, is diagnostic. Although the 12-lead systemprovides a more thorough coverage of ECG function-alities, it is also more costly, both financially and interms of transport time. Therefore, a 3-lead system ischosen for this application.

2.2. Wireless Sensor Networks

Wireless sensor networks (WSN) research is originallymotivated by military applications such as battlefieldsurveillance. As the field slowly matured and technol-ogy rapidly advanced, it has found itself merging intomany of the civilian applications as well, such as en-vironment and habitat monitoring, home automations,traffic control, and more recently healthcare applica-tions. Often equipped with wireless communicationdevices and microcontrollers, distributed autonomous

devices using sensors to cooperatively monitor physi-cal or environmental conditions, such as temperature,sound, vibration, pressure, motion or pollutions, at dif-ferent locations form a wireless sensor network [12].Figure 3 shows three different types of sensor motesthat were used in this research work. They are mostoften referred as motes and come with some sort of apower source.

Developed primarily by the University of California,Berkeley in cooperation with Intel Research, TinyOS isan open-source embedded operating system designedfor wireless sensor networks. Written in NesC pro-gramming language, TinyOS offers a component-basedarchitecture and is able to operate within the severememory constraints posted by sensor networks. Thecopy of TinyOS used in this research is version 1.1.15,released in December of 2005. NesC is a program-ming language designed for applications targeting theTinyOS platform. Again by University of California,Berkeley and Intel Research, it is an extension to the Cprogramming language that is component based as theTinyOS operating system. The most important featureof this programming language is that it produces fairlysmall-sized code to be able to load on to sensor networknodes.

Fig. 3. Sensor motes—Mica2, Mica2Dot, and Telosb.


DOI: 10.1002/wcm


3. Mobile Telemedicine Platform

This section will discuss the functionalities and theconstruction of the hardware mobile platforms. Thesemobile platforms are essentially the wearable devicesthat would be distributed among patients in orderto offer continuous monitoring of the patients’ vitalsigns. As shown in Figure 4, each platform is com-posed of a customized sensor board providing con-nections to a 3-Lead ECG monitoring system, whichis housed on a commercially available TelosB sensormote [21]. While the sensor board gathers useful patientECG data, the sensor mote provides limited processingcapabilities and more importantly wireless communi-cation for transmitting the signals back to the worksta-tion for feature extraction.

3.1. TelosB Mote

The TelosB mote is also referred to as the TmoteSky. Designed at University of California, Berkeley,by TinyOS developers, it is an ultra low power wire-less module intended for sensor networks applications.Regarded as the next-generation mote platform, it of-fers the most on-chip RAM of 10 kB and also thefirst to provide IEEE 802.15.4 Chipcon radio with anintegrated on-board antenna providing up to 125 mof range. Constructed around a TI MSP430 micro-

controller, the TelosB was the ideal choice for thisproject for its on-board ADC peripherals with expan-sion bays, from which the customized sensor board isconnected to.

3.2. Sensor Board

3.2.1. Circuit design

The design of the sensor board is contributed by Har-vard University as part of their ongoing research inproject CodeBlue [13]. As mentioned before, vitalsign monitoring has been a frequently visited topicwith CodeBlue being one of the most successful inacademia. It focuses on the exploration of wireless sen-sor network technology for a range of medical applica-tions and offers a platform that provides protocols fordevice recovery and publish/subscribe multihop rout-ing including query interface for medical monitoring[1]. Although not yet complete at the time of this docu-mentation, its design in Reference [13] for the wearableECG platform has seen much utility as the foundationfor the work conducted in this research.

3.2.2. Layout design

The design of this sensor board may be translated intolayout diagram, which is used for PCB board fabrica-

Fig. 4. Mobile platform.


DOI: 10.1002/wcm


tion, sponsored by PCBExpress. This is a multi-layeredlayout design. The layers are: top silkscreen, top soldermask, top copper, bottom copper, bottom solder mask,and the drill layer. For convenience, the board outlineis part of each layer.

The two layers of coppers are laid onto the other-wise non-conductive board to create the schematic de-sign onto the PCB. Multi-layered PCBs are formed bybonding together separately etched thin boards. De-fined by the solder mask layers, the bare coppers aretypically plated with solder to avoid copper oxidiza-tion, which would make the PCB not solderable. On theother hand, the areas that are not meant to be solderedare covered with solder resist. Lastly, silk-screening isalso called screen printing, where line art and text maybe printed onto the outer surfaces of a PCB. It is oftenused to indicate individual components. The Gerberfile of each layer was sent to the PCBExpress for fab-rication together with a drill file that included all theautomatic drilling information such as hole sizes andmachine coordinates.

3.2.3. Electronic components

A list of components was soldered onto each sensorboard PCB according to the schematic. The completepart list may be found in Table I, where the only cat-egory requiring explanation is Package. The differentstyles of packaging information are the surface mount-ing types that correspond to the PCB layout choices.The complete sensor board is shown in Figure 5 withoutthe TelosB mote.

Fig. 5. Complete sensor board.

The ECG lead extensions from the sensor board arepin-compatible and color coded to standard 3-LeadECG monitoring systems. While there are different fla-vors of physiological chest leads, this system was de-signed to match any 3-Lead ECG Snap Set Leadwiresas shown in Figure 6. The Snap Set may be used to col-lect data by attaching to it the appropriate jellied ECGconductive adhesive electrodes if real people were tobe used for testing purposes. An alternative would beECG signal simulators.

3.3. ECG Signal Simulator

The testing simulator of choice in this project is theModel 430B, 12-lead ECG simulator as shown in

Table I. Sensor board part list.

Quantity Component Package Function

1 INA3 21 8-MSOP Instrumentation amplifier1 OPA4336 16-SSOP Operational amplifier2 100k, 5% 0805 Resisters3 1M, 5% 0805 Resistors3 47k, 5% 0805 Resistors2 2.2M, 5% 0805 Resistors1 3.3k, 5% 0805 Resistors1 4.7k, 5% 0805 Resistors1 806, 5% 0805 Resistors2 0.1 uF 0805 Capacitors2 0.1 uF 1913 Capacitors1 4.7 nF 1206 Capacitors1 1 uF 1210 Capacitors3 J539-ND N/A MCX Jack3 J680-ND N/A MCX Plug3 Wires N/A 20 AWG wires3 A14299-AD N/A Wire crimps1 A26455-ND N/A Connection receptor


DOI: 10.1002/wcm


Fig. 6. 3-Lead ECG snap set leadwires.

Figure 7. This simulator provides a complete PQRSTwaveform at six preset rates (60, 75, 100, 120, 150, and200 BPM) as well as six preset amplitudes (0.1, 0.2, 0.5,1.0, 2.0, and 5.0 mV). It is also capable of generatingsquare waves using its 5 ECG snaps plus 10 bananajacks. This provides a good testing interface even ifthis project will be adapted into a 12-lead monitoringsystem in the future.

Figure 8 shows the connection between 430B ECGsimulator and our designed RF communication boards.The ECG signal collected from 430B can be transmit-ted to a computer (not shown in Figure 8) through theRF board antenna.

4. ECG Data Communications

This section focuses on the communication methodsnecessary to transmit the patient data collected via mo-bile platforms to the feature extraction unit on the work-station. This is a two-step process. The first step in-volves the sensor network communication that takes

Fig. 8. Mobile platform patient simulation.

place between the mobile platforms and the receivingsensor mote connected to the workstation. After thisstep, all of the useful patient data have been collectedand now reside onboard the workstation. The next levelof communication occurs within the workstation envi-ronment, where a MATLAB server is created to transferdata from a Java runtime environment into the MAT-LAB workspace via localhost connection. This is thefinal procedure before sending the patient data for sig-nal processing, which leads to feature extraction.

4.1. Sensor Network Communication

4.1.1. Network configuration

Wireless communication greatly increases the func-tional range and mobility of the system. It is also oneof the advantages that come along the selection sensornetworks as the backbone communication architecture.The sensor mote of choice in this application is theTelosB devices with IEEE 802.15.4-compliant radiocapabilities and a range of 125 m. Other alternativesinclude the Mica2, Mica2Dot motes, although due to

Fig. 7. Model 430B patient simulator.


DOI: 10.1002/wcm


Fig. 9. Wireless sensor network.

their non-standard radio chips, their designs are nowdeprecated in favor of the MicaZ device with standardradio as TelosB does.

These sensor motes operate under the TinyOSenvironment and may be programmed using the NesClanguage. Indeed, the TinyOS system and librariesthemselves are also written in NesC. This style of pro-gram is component-base structured and supports con-currency modeling. NesC applications are built out ofcomponents with bidirectional interfaces and are thenlinked together to form an executable program. Theconcurrency model provides two threads of execution:tasks and hardware even handlers. Tasks are scheduledfunctions that run to completion without the abilityto preempt one another. Hardware event handlers areresponses to desired hardware interrupts that preemptexecutions of tasks and other hardware even handlers.

The wireless sensor network is composed of twogroups of devices classified based on their operations.As seen in Figure 9, the two groups are mobile plat-forms and the receiving station. The mobile platforms,as previously explained, are meant for patient data col-lection and are distributed to the patients as wearabledevices. The quantity of mobile platforms in each sys-tem may vary depending on the number of patients’needs to be observed. There is, however, only one re-ceiving station in each system setup. Connected to theworkstation via an USB port, the receiving station ismeant for data gathering and actively communicateswith each of the mobile platforms in use.

4.1.2. Sensor communication software

The software used to govern the sensor network com-munication and displaying the received patient data onthe workstation is based on a program called VitalDust

Plus. VitalDust Plus is developed by a group of Ph.D.students at Harvard University led by Dr Matt Welsh.This software is essentially a stripped down version ofthe CodeBlue software that provides a simple demon-stration of its wireless pulse oximeter and wireless ECKdevices. The software has two parts, the TinyOS soft-ware for the mobile platforms to sample and transmitvital sign data over the radio, and a Java GUI applica-tion to display the vital signs a graphical form.

VitalDust Plus made several functional additions tothe Java applications. The most notable modificationsare the inclusions of MATLAB support and the abilityto select data, at run time, from only the desired patientfor feature extraction. Some of the unused features arealso removed from the original graphical user interface.Figure 10 shows a screen shot of the software while it isreceiving patient data from two separate mobile plat-forms: mote30 and mote40. The patient data field isdisplaying the ECG waveform associated with the se-lected mobile platform. Only data from the currentlyselected mobile platform are sent to MATLAB for sig-nal processing. The link quality field shows the qualityof the wireless signal also associated with the selectedmobile platform.

4.2. MATLAB Server

Although MATLAB has proven itself to be a very pow-erful instrument in both academia and industry, it doesnot provide command line support for its functions andlibraries outside the MATLAB working environment.This is particularly cumbersome for its intended ap-plications in this research. MATLAB and its wavelettoolbox provide a good option for the desired signalwavelet analysis and feature extraction, but the patientdata are passed in automatically via a Java applicationin a complete separate working environment. WhileMATLAB does provide Java Virtual Machine support,it is not possible the other way around to access MAT-LAB functions from a Java program outside. To main-tain the real-time behavior of this application, the pa-tient data must be passed into the MATLAB workspacepromptly for signal processing.

The solution to the above problems is to set up aMATLAB server establishing a connection to the lo-calhost that enables communication within the work-station. A number of additional files are required tomake this work classified into the server side and theclient side. The MATLAB server is based on a smallapplication named MatlabControl.java developed byKamin Whitehouse during his studies at University ofCalifornia, Berkeley. This is a Java program intended to


DOI: 10.1002/wcm


Fig. 10. VitalDust Plus.

access MATLAB commands while running inside theMATLAB working environment. This is made possi-ble by MATLAB’s support for the Java Virtual Environ-ment and the abilities to execute normal Java programs.

The MATLAB server is based on the MatlabControlfile. It establishes a localhost connection and awaitscommunication from the outside programs. Upon re-ceiving messages, it either redirects them to the appro-priate MATLAB functions via MatlabControl.java, orresponds with a predefined solution back to the await-ing clients. One of the problems that exists with runninga Java program inside the MATLAB environment is thefact that MATLAB provides only one single thread,therefore the termination of any Java application initi-ated from inside MATLAB would also exit MATLABas well.

The client side of program is incorporated into theFlavor RIT application by reading patient data fromFlavor RIT and communicates it to the MATLABserver via the established localhost connection. How-ever, due to the continuous input of patient data fromthe mobile platforms, it is impossible to send all of

them at the same time, especially during times whenthere are more than one connected mobile platform.The design choice was to only send in data associatedwith the currently selected in Flavor RIT for waveletanalysis after every 600 packets have been collected.This provides a meaningful mediation for data process-ing and data displaying. A sample extraction result isshown in Figure 11.

5. ECG Feature Extraction

Feature extraction is a commonly used term in imageprocessing and pattern recognition. It is a form of di-mensionality reduction that locates points of interestfrom a multidimensional space. In the scope of thisresearch, feature extraction is conducted by applyingwavelet analysis techniques to patient data, thus pro-viding ECG characteristic-point detection capabilities.Since most recently published detectors are based onstandard database libraries, this real-time application isan attempt to expand the horizons of current research


DOI: 10.1002/wcm


Fig. 11. Data displayed every 600 packets.

efforts. It also offers a significant function extension toexisting vital sign monitoring systems and brings themone step closer to medical care realization.

5.1. Analysis Overview

The analysis of ECG is widely used for diagnosingmany cardiac diseases, the main cause of mortality indeveloped countries [9]. The automatic detection ofECG wave is an important topic, especially for ex-tended recordings, because it provides many clinical in-sights which can be derived from the information foundin the intervals and amplitudes defined by the signif-icant points. The performance of such automatic sys-tems relies heavily on the accuracy and reliability in thedetection of the QRS complex, which is necessary todetermine the heart rate, and as reference for beat align-ments. The QRS complex is the most characteristicwaveform of the signal with higher amplitudes. It maybe used as references for the detection of other waves,such as the P and T complexes, which are also usefulat times. The feature extraction methods applied in thisresearch focuses on the detection of the QRS complexand characteristic points in addition to attempting tolocate the associated P and T waves if there are any.

Wavelet transform is perceived as a very promisingtechnique for this type of applications because it is lo-calized in both the frequency and time domain. It may

be used to distinguish ECG waves from serious noise,artifacts, and baseline drift [4]. Wavelet transformationrepresents the temporal features of a signal at differentresolution, providing better analysis of ECG signals,which is characterized by cyclic occurring patterns atdifferent frequencies. The wavelet transformation is notdifficult to apply as a mathematical tool for decompos-ing signals. The real difficulty comes at choosing amother wavelet that optimally fits the signal dependingon the application and the signal itself.

It is necessary to consider a few important charac-teristics when selecting the optimal mother wavelet.To reconstruct the signal from the wavelet decomposi-tions and to preserve the energy under the transforma-tion are important [8]. Symmetry is another importantcharacteristic in avoiding a drift of the information [8].Although there are many different proposed waveletanalysis algorithms in academia, most of them have onecommonality, which is the selection of discrete wavelettransform over continuous wavelet transforms.

Discrete wavelet transform has its natural advantageswhen applied towards ECG analysis. Conventionally,ECG feature extraction is preceded by a bandpass or amatched filter to suppress the P and T waves and noisesbefore sending the signal for characteristic detection.By using discrete wavelet transform, frequency domainfiltering is implicitly performed so that the system be-comes robust and the direct application over raw ECG


DOI: 10.1002/wcm


Fig. 12. Filter banks analysis.

signals is allowed [9] Again, this is made possible dueto the nature of the discrete wavelet transform. Discretewavelet transform is also referred to as decompositionby wavelet filter banks as shown in Figure 12.

It uses two filtering banks, a low-pass filter and ahigh-pass filter to decompose the signals into its dif-ferent scales. This process may be iterative into manylevels as shown in Figure 13. It can be seen that the dis-crete wavelet transform of a signal x[n] is calculated bypassing it through a series of filter banks, of which theresult may be interpreted as the convolution of the sig-nal itself with the corresponding impulse responses asshown in Equations (1) and (2). From the equations, gis the impulse response of a low-pass filter while h isthe high-pass filter. The output from the high-pass fil-ter is regarded as the detail coefficients and the outputfrom the low-pass filter is regarded as the approxima-tion coefficients. The filters are known as a quadraturemirror filter where a filter bank splits an input signalinto two bands which are usually then sub-sampled bya factor of 2 [12].

y [n] = (x × g) [n] =∞∑

k=−∞x [k] · g [n − k] (1)

Fig. 13. Haar wavelet.

y [n] = (x × h) [n] =∞∑

k=−∞x [k] · h [n − k] (2)

To compensate for the fact that the amount of datawould double after each filtering stage due to the dualbank architecture, the filter outputs are down-sampledby two. However, according to Nyquist’s rule, this mea-sure does not compromise the original level of accu-racy. In this case, Equations (1) and (2) may be rewrit-ten as Equations (3) and (4).

ylow [n] =∞∑

k=−∞x [k] · g [2 · n − k] (3)

yhigh [n] =∞∑

k=−∞x [k] · h [2 · n − k] (4)

Equation (5) shows the overall discrete wavelet trans-form using the down-sampling operator ↓.

(y ↓ k) [n] = y [k · n] (5)

To be more precise, it may be separated into Equations(6) and (7) below.

ylow = (x × g) ↓ 2 (6)

yhigh = (x × h) ↓ 2 (7)

5.2. Algorithms

Due to the nature of wavelet transformations the orig-inal signal is decomposed into wavelets that are dila-tions of a mother wavelet by a scale factor. The key toa successful wavelet analysis is by selecting the opti-mal mother wavelet function that fits the signals takinginto consideration the application and the signal itself.This is no simple task. Although wavelet analysis is a


DOI: 10.1002/wcm


rather new method from an historical point of view, ithas spurred tremendous interest and, as a result, dozensof families of wavelets were born. This section will in-troduce a few of the basic and prominent ones.

The earliest mother wavelet recording refers backall the way to the early 1900s when Alfred Haar firstintroduced the Haar wavelet. As shown in Figure 13,the Haar wavelet is the first and simplest wavelet. It isdiscontinuous and therefore not differentiable. Resem-bling a step function, it is indeed a special case of theDaubechies wavelet and is also known as db1, wherethe number represents the order.

The Daubechies wavelets are invented by IngridDaubechies, a top wavelet researcher who made prac-tical the fields of discrete wavelet analysis. TheDaubechies family wavelets are represented as dbN,where N is the order. Figure 14 shows the next ninemembers of the family skipping db1, which is the Haarwavelet.

The order number of the Daubechies wavelet mayextend to a much higher range. This is a family oforthogonal wavelets defining a discrete wavelet trans-form characterized by a maximal number of vanishingmoments from some given support. It provides solu-tions for applications such as self-similarity propertiesof a signal or fractal programs, signal discontinuities,etc. Daubechies went on to propose another waveletfamily marked by its symmetrical attribute and are de-rived from the db family. This new family of waveletsis referred as the symlets and is shown in Figure 15.The two families of wavelets exhibit very similarproperties.

There are also a wide range of other wavelet fam-ilies such as the Biorthogonal family that uses two

wavelets, one for decomposition and other for recon-struction; the Morlet family that has no scaling func-tions; the Meyer family whose wavelet and scalingfunctions are defined in the frequency domain; as wellas many other real and complex wavelets. Again the keyis to find one that would best represent the signals athand.

One of the most notable feature extraction algo-rithms for ECG characteristic-point detection is pre-sented by Li et al. [4] in 1995 at the Biomedical Engi-neering Institute of Xi’an Jiaotong University. This of-ten referenced publication proposed a multi-scale QRSdetector including a method for detecting monophaticP and T waves. The algorithm utilizes the discretewavelet transform advantage and may be applied overraw ECG signals without any pre-filtering. A quadraticspline wavelet with compact support and one vanishingmoment proposed by Mallat and Zhong in Reference[18] was used as the mother wavelet.

Most of the energy of ECG signals is between 21

and 25 [9]. For scales higher than 24, the energy ofthe QRS is very low [9]. The P and T waves have sig-nificant components at scale 25 although the influenceof baseline wandering is important [9]. This result isbased on the equivalent responses from Figure 16 andaccording to the spectrum of the ECG signal waves byThakor et al. in Reference [19].

This algorithm uses the information of local max-ima, minima, and zero crossing at different scales. Itassociates each change in the signal with a line of max-ima or minima crossing the scales as shown in Figure17. The algorithm follows four steps: detection of QRScomplexes, detection and identification of the QRS in-dividual waves and boundaries, T wave detection and

Fig. 14. Daubechies wavelets.


DOI: 10.1002/wcm


Fig. 15. Symlets wavelets.

delineation, and lastly P wave detection and de-lineation. However, only the QRS detector wasvalidated.

Even though this algorithm was not actually imple-mented, it contributed many fundamental theories inthe field of ECG characteristic-point feature extrac-tions. The algorithm’s complexity posed challenges inassuring the performance of the proposed system inreal-time settings. Its implementation of a non-industrystandard mother wavelet also made its implementa-tion more worrying for conventional situations. Analternative method was selected to ensure the perfor-mance and easy implementation for this proposed sys-tem. The selected algorithm is presented in the nextsection.

Fig. 16. Equivalent frequency responses of the DWT at scales2k, k = 1, . . . , 5 for 250 Hz sampling rate.

5.3. Implementation

The implemented feature extraction algorithm is basedon the work presented by Mahmoodabadi et al. in Ref-erence [7]. While there are no absolute guidelines inselecting a wavelet family, it is of utmost importancethat the wavelet function closely matches the signalto be processed [20]. As described in the last section,there are many flavors of wavelet families available.Although the Haar wavelet provides the benefits ofsimplicity, it does not take into consideration the finerdetails of a signal. The Daubechies Wavelets are con-ceptually more complex than the Haar wavelet and aresimilar in shapes to QRS complex of ECG waves. Theirenergy spectrum is also concentrated around the lowfrequencies, making them the mother wavelet of choicein this application.

The specific Daubechies wavelet use in this imple-mentation is db6 as suggested by Mahmoodabadi. Fig-ure 18 shows a 5-level signal decomposition of a sam-ple ECG waveform using this wavelet. It also includesa comparison between the original ECG signal and thereconstructed ECG signal. This is important becauseone of the key criteria of a good mother wavelet is itsability to fully reconstruct the signal from the waveletdecompositions.

The high frequency components of the ECG signaldecreases as lower details are removed from the origi-nal signal. As the lower details are removed, signal be-comes smoother and the noises on the T and P wavesdisappear since noises are marked high frequency com-ponents picked up along the ways of transmission. Thisis the contribution of the discrete wavelet transform


DOI: 10.1002/wcm


Fig. 17. WT at the first five scales of ECG-like simulated waves.

where noise filtration is performed implicitly. This isexplained by the ECG signal frequency distribution,which is shown in Figure 19.

The feature extraction algorithm follows the follow-ing steps: R wave detection, Q and S wave detection,zero-level detection, and lastly P and T wave detec-

tion. There are actually four algorithms, each focusingon one certain feature of the ECG signal. The result ofthe previous detections may be used as references inthe later detections. Although sequential in nature, allof the algorithms are applied directly at one run overthe entire digitized ECG signals collected using the

Fig. 18. 5-Level decomposition using DB6.


DOI: 10.1002/wcm


Fig. 19. Normal ECG signal and frequency distribution.

Fig. 20. Sample feature extraction.


DOI: 10.1002/wcm


Fig. 21. Real-time feature extractions.

mobile platforms. The detailed algorithms of the fea-ture detection methods are listed below.

5.3.1. R peak detection

The detection of the R peaks is the first step of featureextraction. The patient data are broken into segmentsof 600 points and only one segment is analyzed at atime. The R peaks have the largest amplitudes amongall the waves, making them the easiest to detect andgood reference points for future detections. The sig-nal was processed using the db6 wavelet up to eightlevels. However, for the detection of R peaks, only de-tails up to level 25 were kept and all the rest removed.This procedure removed lower frequencies consider-ing QRS waves have comparatively higher frequencythan other waves. The attained data are then squared tostress the signal. A threshold equalling 30% of the max-imum value is sub-sequentially applied to set a practicallower limit to help to remove the unrelated noisy peaks.At this point, the data set is ready for peak detectionthrough a very simple search algorithm that producesvery accurate results.

5.3.2. QS detection

The detection of Q and S peaks is associated directlywith the detection of R peaks. Q and S peaks occursabout the R peaks within 0.1 s. Therefore this detectionalgorithm requires the results from the previous part forsetting up windows of interest. Only details up to level24 are used for searching of the extermum points abouteach R peak formally detected. The point precedingthe R peak denotes the Q peak and the point followingthe R peak denotes the S peak. ‘A normal QRS com-plex indicates that the electrical impulse has progressednormally from the bundle of His to the Purkinje net-work through the right and left bundle branches andthat normal depolarization of the right and ventricleshave occurred’ [7].

5.3.3. Zero-level detection

The zero level of a recording ECG is the point wherethere is no current flowing around the heart. Thispoint is difficult to attain because there are manystray currents existing in the body resulting from skin


DOI: 10.1002/wcm


potentials and differences in ionic concentrations in dif-ferent parts of the body. Conventionally researchers areconditioned to consider the TP segment as the zero-point reference level. However, the point right after theend of the QRS complex marks the real zero-level po-tential. This is known as the J point, where even thecurrent of injury disappears [11]. There are two zero-level points, one before the Q peak and one after the Speak. Decomposition details of up to level 24 and theapproximation level 28 are used for J point detections.Although not part of the actual feature extraction pro-cedure, these points are used as reference points in thedetection of the PT waves.

5.3.4. PT wave detection

Although the work in this research focuses on the detec-tion of the QRS complexes, an effort has been made toprovide some information regarding the P and T wavesthat may be present in the ECG signals. PT waves pro-vide meaning supplement information to the QRS com-plexes in detection of cardiac diseases. The detectionof the PT waves in this algorithm requires knowledgeof the J points while using decomposed signals of onlythe level 28 approximation signals. P wave is detectedbefore the first J point preceding the R peaks, while Twave is detected after the second J point following theR peaks. The J points may be considered to be the onsetand offset points of the waves, respectively.

5.3.5. Results

All four of the above algorithms are applied at oncein real time to the collected patient data via the mobileplatforms. The patient data are segmented by FlavorRIT and the MATLAB server into 600-point packages.Each package is sent to the MATLAB workspace atone time for signal extraction. This process is repeatedpacket after packet producing the results in a MAT-LAB figure. Some of the sample results are shown be-fore. Figure 20 shows the feature extraction applied toa software generated sample ECG data set while Figure21 shows the feature extraction result applied over thereal-time data collected via the mobile platforms.

The red cross sign denotes the R peaks, and as seen,it is located 100% of the time. The Q and S peaks aredenoted, respectively, by black plus sign and black mul-tiplication sign. Although with occasion miss place-ments (1 in Figure 21), they still provide very accuratedetections. The black diamonds and black squares de-note the P and T waves, they show the approximationof the PT waves that were present in the ECG signals.

6. Conclusions

The objective of this research was to take advantage ofthe modern day technology and create a tele-cardiologysensor network for remote ECG monitoring purposes.Our system is composed of two major components.Based on wireless sensor network technology, there arewearable mobile platforms distributed to the patientsof concern. These mobile platforms are responsible forgathering patient vital sign using a 3-Lead ECG moni-toring system. The gathered data are transmitted wire-lessly over radio to the receiving station connected to aworkstation where the data are processed. The secondpart of the system is based on wavelet analysis theories.This takes place onboard the workstation where patientdata are gathered. Feature extraction techniques are ap-plied to the patient data and the characteristic points ofinterests extracted. These data provide meaningful in-formation for the diagnosis of possible cardiovasculardiseases. This is especially useful for extended record-ings of ECG signals where human processing is notonly time consuming, but also error prone.

A future expansion possibility is studied for loca-tion tracking. This is an important expansion becausethe original intention of this system is to decrease theamount of time required for medical response to pa-tients in need. By having the exactly patient locationsin hand, it is possible to further reduce the responsetime.

Acknowledgement

This research has been partially supported by Sensor-con, Inc. Biomedical Sensor project (2004–2006).

References

1. Shnayder V, Chen B, Lorincz K, Fulford-Jones T, Welsh M.Sensor networks for medical care. Technical Report TR-08-05.Division of Engineering and Applied Sciences, Harvard Univer-sity, 2005.

2. Lorincz K, Welsh M. MoteTrack: a robust, decentralized ap-proach to RF-based location tracking. In Proceedings of theInternational Workshop on Location and Context-Awareness(LoCA 2005) at Pervasive 2005, May 2005.

3. Karlof C, Sastry N, Wagner D. TinySec: a link layer security ar-chitecture for wireless sensor networks. SenSys’04, 3–5 Novem-ber 2004. Baltimore, Maryland, USA, 2004.

4. Li C, Zheng C, Tai C. Detection of ECG characteristic pointsusing wavelet transforms. IEEE Transactions on Biomedical En-gineering, Vol. 42, No. 1, January 1995. Biomedical Engineer-ing Institute of Xi’an Jiaotong University, Xi’an, Shaanxi, P. R.China, 1993.


DOI: 10.1002/wcm


5. Senhadji L, Carrault G, Bellanger JJ, Passariello G. Comparingwavelet transforms for recognizing cardiac patterns. IEEEEngineering in Medicine and Biology. Laboratoire Traitementdu Signal et de l’Image (INSERM) Universite de Rennes I.Rennes, France. Grupo de Bioingeneria y Biofisica Aplicada,Universidad Simon Bolivar. Caracas, Venezuela, 1995.

6. Daqrouq K. ECG baseline wandering reduction using discretewavelet transform. Asian Journal of Information Technology2005; 4(11): 989–995. Department of Electronics and Commu-nication Engineering, University of The Philadelphia, Amman,Jordan.

7. Mahmoodabadi SZ Ahmadian A, Abolhasani MD. ECG featureextraction using Daubechies wavelets. Proceedings of the FifthIASTED International Conference, Visualization, Imaging, andImage Processing. Benidorm, Spain. 7–9 September 2005.

8. Castro B, Kogan D, Geva A. ECG feature extraction using op-timal mother wavelet. The 21st IEEE Convention of the Elec-trical and Electronic Engineers in Israel, 2000. Department ofElectrical and Computer Engineering, Ben-Gurion University,Beer-Sheva, Israel, 2000.

9. Martinez JP, Almeida R, Olmos S, Rocha AP, Laguna P.A wavelet-based ECG delineator: evaluation on standarddatabases. IEEE Transactions on Biomedical Engineering 2004;4(4).

10. Chan YT. Wavelet Basics. Kluwer Academic Publishers: Boston,1995.

11. Guyton AC, Hall J. Textbook of Medical Physiology (10th edn).W.B. Saunders Company: Philadelphia, PA, 2000.

12. Wikipedia. the free encyclopedia. http://www.wikipedia.com.13. Welsh M, Chen B. CodeBlue: Wireless Sensor Networks for

Medical Care. Division of Engineering and Applied Sciences,Harvard University, 2006. http://www.eecs.harvard.edu/∼mdw/proj/codeblue/.

14. Meyer Y, Ryan RD. Wavelets: algorithms & applications. TheSociety for Industrial and Applied Mathematics, 1994.

15. Wavelet Toolbox Documentation, the MathWorks, MATLABToolbox. http://www.mathworks.com/access/helpdesk/help/toolbox/wavelet/wavelet.html?/access/helpdesk/help/toolbox/wavelet/ch01 i11.html.

16. MATLAB Product Page, http://www.mathworks.com/products/matlab/.

17. Tmote Sky Datasheet. http://www.moteiv.com.18. Mallat S, Zhong S. Characterization of signals from multiscale

edge. IEEE Transactions Pattern Analysis and Machine Intelli-gence. 1992; 14: 710–732.

19. Thakor NV, Webster JG, Tompkins WJ. Estimation of QRS com-plex power spectrum for design of a QRS filter. IEEE Transac-tions on Biomedical Engineering 1984; BME-31: 702–706.

20. Grap A. An introduction to wavelets. IEEE Computer Scienceand Engineering 1995; 2(2): 50–61.

21. TesloB mote: see http://www.xbow.com for more details.22. Xiao Y, Shen X, Sun B, Cai L. Security and privacy in RFID and

applications in telemedicine. IEEE Communications MagazineSpecial Issue on Wireless Technology Advances and Challengesfor Telemedicine, April 2006; 64–72.

23. Hu F, Kumar S, Xiao Y. Towards a secure, RFID/sensor basedtele-cardiology system. 2007 IEEE Consumer Communica-tions and Networking Conference—Special Sessions Track onTelemedicine, Las Vegus, January 2007.

24. Xiao Y, Hu F. Wireless Telemedicine and M-Heath. 2007 IEEEConsumer Communications and Networking Conference—Special Sessions Track on Telemedicine, Las Vegus, January2007.

25. Xiao Y, Takahashi D, Hu F. Telemedicine usage and potentials.IEEE Wireless Communications and Networks Conference 2007,March 2007.

Authors’ Biographies

Fei Hu received his B.S. and M.S. de-grees from Shanghai Tongji University(China) in 1993 and 1996, respectively.He received his Ph.D. degree from theDepartment of Electrical and ComputerEngineering at Clarkson University in2002. His Ph.D. research was on high-performance transmission issues in wire-less networks. He is currently an Assis-

tant Professor in the Computer Engineering Department atRIT, New York. He served as a Senior Networking Engineerat the Shanghai Networking Lab and Shanghai Lucent, Inc.from 1996 to 1999, where he worked on several large projectson high-performance networks. Dr Hu is a full Sigmaxi mem-ber, a member of the IEEE, and an IEEE chapter officer. Hisresearch interests are in ad hoc sensor networks, 3G wirelessand mobile networks, and network security. His research hasbeen supported by NSF, Cisco, Lockheed Martin, Sprint, andso on.

Meng Jiang is currently with Intel inU.S.A. He obtained his B.S. and M.S.(dual degrees) from Computer Engineer-ing Department at RIT (Rochester, NY,U.S.A) in 2006. His research interests arebiomedical chip design and wireless in-tegration.

Yang Xiao worked at Micro Linear asa MAC architect involved in the IEEE802.11 standard enhancement work be-fore he joined University of Memphis in2002. Dr Xiao is currently with Depart-ment of Computer Science of Univer-sity of Alabama in 2006. He was a vot-ing member of the IEEE 802.11 Work-ing Group from 2001 to 2004, and is an

IEEE senior member. He currently serves as Editor-in-Chieffor International Journal of Security and Networks and forInternational Journal of Sensor Networks. He serves as As-sociate Editor or on editorial board for five referred journals.He served as a Guest Editor for eight journal special issues.He serves as a panelist for NSF, and a member of CanadaFoundation for Innovation (CFI)’s Telecommunications ex-pert committee. His research areas include wireless networks,mobile computing, and network security. He has publishedmore than 70 journal papers with more than 40 papers pub-lished in various IEEE journals. He has edited/co-edited 10books on wireless networks and security.


DOI: 10.1002/wcm

low-cost wireless sensor networks for remote cardiac patients monitoring applications

Documents